DeepSeek has released DeepSeek-V4 Preview as open-source, introducing two models: V4-Pro (1.6T total / 49B active parameters) and V4-Flash (284B total / 13B active parameters). Both models feature a 1M token context window as the new default, powered by a novel attention mechanism combining token-wise compression and DeepSeek Sparse Attention (DSA). V4-Pro claims open-source SOTA on agentic coding benchmarks and leads open models in math/STEM/coding, rivaling top closed-source models. V4-Flash offers near-V4-Pro reasoning at faster speeds and lower cost. The API is available today with OpenAI ChatCompletions and Anthropic API compatibility. Existing deepseek-chat and deepseek-reasoner endpoints will be retired on July 24, 2026.
Table of contents
DeepSeek-V4-Pro DeepSeek-V4-Flash Structural Innovation & Ultra-High Context Efficiency Dedicated Optimizations for Agent Capabilities API is Available Today! Sort: